Lots of Reasons to Worry about AI Infra, But Customer Concentration Cannot be Helped

There are lots of reasons for investors to worry about the artificial intelligence infrastructure business. After all, firms generally are committing capital to infrastructure faster than the economic value of AI is becoming visible. 

The investment question is therefore whether AI demand will grow fast enough, and profitably enough, to absorb the enormous quantity of infrastructure being financed today.

Concern	What investors worry about	Why AI infrastructure is particularly exposed	Potential consequence
Overbuilding / excess capacity	Infrastructure is being built on expectations of enormous future AI demand that may not materialize at the projected rate	Data centers and GPUs are highly capital-intensive and difficult to redeploy	Falling utilization, lower rental rates and impaired asset values
Return on invested capital	AI revenue may grow rapidly but still fail to produce returns commensurate with the enormous capital investment	A $1 billion data center requires substantial utilization for years to earn an attractive return	Lower ROIC and pressure on hyperscaler valuations
Rapid technology obsolescence	Today's expensive GPU cluster may become economically inferior before its expected useful life ends	AI accelerators are advancing unusually rapidly	Accelerated depreciation and residual-value losses
Falling AI compute prices	Competition and increasingly efficient models could cause the price of compute to fall faster than infrastructure costs	Compute is becoming increasingly commoditized	Infrastructure operators could experience margin compression
Financing leverage	The industry increasingly needs debt, leases, project finance, guarantees and SPVs to fund expansion	Capital requirements have grown faster than many companies' internally generated cash	Credit deterioration and refinancing risk
Off-balance-sheet exposure	Some economic obligations may be less obvious than conventional corporate debt	Guarantees, leases, joint ventures and residual-value guarantees complicate analysis	Investors discover greater effective leverage during a downturn
Customer concentration	A data center may effectively depend on one or two giant AI customers	New facilities are enormous and frequently built around anchor tenants	Default or spending cuts by one customer can undermine project economics
Circular financing / ecosystem dependency	Suppliers, customers and financiers may increasingly finance one another	Nvidia, hyperscalers, AI labs, data-center developers and private capital are becoming financially intertwined	Difficulty determining how much demand is truly independent
Power constraints	Infrastructure may be valuable but unusable because electricity cannot be delivered on schedule	AI facilities require enormous amounts of concentrated power	Delays, stranded land/capacity and higher costs
Permitting and community opposition	Local governments or residents may delay or block projects	AI data centers have unusually large power, water, land and noise footprints	Construction delays and unexpected costs
Grid / energy-price risk	Electricity may become substantially more expensive as AI competes for scarce generation	Power is becoming a major component of AI operating costs	Lower data-center margins and higher customer prices
Demand concentration in a few companies	Much of the spending ultimately depends on a handful of hyperscalers and frontier-model companies continuing to spend aggressively	Microsoft, Amazon, Google, Meta, Oracle and a small number of AI labs dominate demand	A capex slowdown could propagate rapidly through the supply chain
AI monetization uncertainty	AI usage may grow enormously without generating enough incremental profits to justify infrastructure spending	Consumers and businesses increasingly expect AI to become cheap or bundled into existing products	Revenue growth can lag infrastructure investment
Macro/interest-rate sensitivity	Higher rates make long-duration infrastructure investments less attractive	Data centers have large upfront costs and long payback periods	Lower valuations and more expensive project financing
Regulatory/geopolitical risk	Export controls, AI regulation, energy policy or restrictions on data centers could change demand	AI infrastructure is concentrated in a relatively small number of countries and suppliers	Assets may become stranded or economically less valuable

But one of the concerns is almost unavoidable. Standard business strategy is to diversify customer bases so that no change at any single customer account will imperil the business overall. But that seems virtually impossible in the AI infrastructure business, especially the “compute as a service” segment. 

The reason is that there simply are very few major buyers. 

Computing Product Category	Dominant Buyer Segment	Market Dynamics & Demand Concentration Drivers
High-Performance AI Accelerators (GPUs/TPUs)	Microsoft, Meta, Google, Amazon (AWS)	Hyperscalers routinely absorb upwards of 40-50% of total advanced enterprise chip allocations, deploying hundreds of thousands of specialized accelerators per cluster to train and serve frontier models.
Custom AI Silicon & ASICs	Google (TPUs), Amazon (Trainium/Inferentia), Meta (MTIA), Microsoft (Maia)	Demand is entirely insourced and consolidated among the top cloud operators designing proprietary chips to reduce dependency on merchant silicon and optimize workload economics.
Enterprise Server Racks & Motherboards	Hyperscale Data Center Operators & Large Cloud Providers	Traditional enterprise server buyers are bypassed by custom Open Compute Project (OCP) specs ordered at massive scale by a handful of operators, leaving original design manufacturers (ODMs) highly concentrated.
High-Bandwidth Memory (HBM)	Major GPU Makers fulfilling Hyperscaler Orders	Production capacity for advanced multi-layer memory stacks is heavily pre-committed to fulfill multi-billion dollar buildouts driven by the major cloud titans.
Advanced Data Center Networking (InfiniBand / High-Speed Ethernet Switches)	Microsoft, Meta, Google, Amazon, Oracle	Ultra-low latency fabric requirements limit early-stage adoption of cutting-edge networking gear to the major cloud providers scaling distributed GPU training fabrics.

But that is not unique to the AI infrastructure business. Some infrastructure markets are inherently oligopsonistic (there may be many suppliers, but only a handful of economically viable buyers).

TSMC doesn't have the option of saying, "If Apple doesn't order this next-generation process, we'll simply find 100 other customers." The universe of customers capable of economically using a 2-nanometer-class process is tiny: Apple, Nvidia, AMD, Qualcomm, Broadcom, MediaTek and a handful of others.

Some other industries also have very-concentrated buyers. In the commercial aerospace business, tier-one avionics and structural component suppliers sell almost exclusively to a duopoly of global aircraft manufacturers: Boeing and Airbus.

In the defense contracting industry, specialized aerospace, radar, and cybersecurity firms rely almost entirely on a single primary buyer: the U.S. Department of Defense or allied national governments.

In the advanced semiconductor equipment industry, companies producing critical lithography systems sell to a handful of chip fabrication giants such as TSMC, Samsung, and Intel.

In railway rolling stock, manufacturers of heavy locomotives and specialized railcars interface with heavily consolidated markets dominated by national freight networks or state-run transit authorities.

Market	Typical number of economically meaningful buyers	Why buyers are so few	Examples of suppliers
AI hyperscale data centers	~5–10	Enormous power, capital and computing requirements mean only hyperscalers and a few AI companies can consume large campuses	Nvidia, Credo, Arista Networks, Cisco, HPE
Advanced semiconductor fabs	~5–10 major customers	Leading-edge chips require enormous volumes and only a few companies design them	TSMC, Samsung, Intel Foundry
Extreme-ultraviolet lithography	Essentially 2–3	Only a handful of semiconductor manufacturers need leading-edge EUV equipment	ASML
Commercial aircraft	~50–100 significant airlines globally, but much smaller number of major customers	Aircraft cost hundreds of millions of dollars and fleet procurement is concentrated	Boeing, Airbus
Large commercial jet engines	~10–20 major airline/airframe customers	Few aircraft platforms and three major engine manufacturers	GE Aerospace, RTX/Pratt & Whitney, Rolls-Royce
Cruise ships	~10–20 major cruise operators	Extremely expensive specialized vessels; only major cruise companies can order them	Meyer Werft, Fincantieri, Chantiers
Nuclear reactors	Very few utilities/developers	Gigawatt-scale projects cost billions and require specialized sites, regulation and financing	GE Vernova, Westinghouse, EDF, KHNP
Military aircraft	A handful of governments	National-security requirements restrict buyers; export markets are politically constrained	Lockheed Martin, Boeing, RTX, Northrop Grumman
Military satellites / launch systems	Very few governments	Classified requirements and national-security restrictions sharply constrain customers	Lockheed Martin, Northrop, Boeing, SpaceX
Large LNG projects / LNG carriers	Relatively few global buyers	Huge projects require long-term contracts and specialized infrastructure	Cheniere, QatarEnergy, shipyards
Electricity generation equipment	Hundreds of utilities, but few very large buyers	Large turbines and grid equipment are bought in relatively infrequent, lumpy projects	GE Vernova, Siemens Energy, Mitsubishi Heavy
High-end semiconductor packaging	Relatively few hyperscalers/chip designers	Advanced packaging is expensive and tied to particular chip architectures	TSMC, ASE, Amkor
Mining equipment for ultra-large mines	A few dozen major mining companies	Huge trucks, shovels and autonomous systems are economically viable mainly at enormous mines	Caterpillar, Komatsu, Epiroc
Offshore oil platforms / FPSOs	A few dozen oil companies	Projects cost billions and are geographically and technically specialized	SBM Offshore, MODEC, Samsung Heavy
High-speed rail equipment	Few national/state buyers	Large projects are government-controlled and geographically constrained	Alstom, Siemens Mobility, CRRC
Container ships	Many shipping companies, but highly concentrated among largest carriers	Very large vessels require substantial capital and are ordered in batches	Hyundai, Hanwha Ocean, CSSC

There are many reasons investors can worry about the health of the AI infrastructure business. But customer concentration does not seem a concern that can realistically be avoided. Some industries are just like that: there are few potential buyers.

If You Want to Conserve Water, Agriculture is Where One Must Look

“For every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California,” OpenAI CEO Sam Altman says. That estimate also likely uses a low figure for ChatGPT water consumption and a higher estimate for almond production. 

Still, it is reasonable to state that an almond takes thousands to tens of thousands of times more water to produce than a single ordinary AI query. 

Comparison	Water
California almond — total water footprint	~12 liters
California almond — freshwater use	~5.8 to 6 L
ChatGPT query — OpenAI/Altman estimate	~0.32 mL
12 L ÷ 0.32 mL	~37,500 queries
6 L ÷ 0.32 mL	~18,750 queries

Other studies, though, do tend to validate the water usage required as between five liters per kilogram of almonds to 13,000 liters of water per kilogram of almonds. There are other methodological issues including how one calculates water consumption used to produce electricity, for example. 

And some of the water used to produce almonds comes from rainfall, which some might not consider a draw on water supplies, though others would likely counter that such water might be used in other ways (to grow different crops; recharge aquifers, flow to rivers and so forth). 

Study/source	Subject	Water-use estimate	What it tells us
Fulton, Norton & Shilling, 2019, Ecological Indicators	California almonds	10,240 L/kg, or ~12 L/almond	Comprehensive water-footprint accounting: blue + green + gray water. (DOI)
Mekonnen & Hoekstra, 2011, Hydrology and Earth System Sciences	Global crop water footprints	California almonds: ~12,984 L/kg in their underlying dataset	Establishes the influential global water-footprint methodology and distinguishes green, blue and gray water. (HESS)
Marvinney & Kendall, 2021, International Journal of Life Cycle Assessment	California almonds	~4,820 L freshwater/kg; ~4,540–5,150 L depending on region	A more direct freshwater-use/LCA measure, excluding rainfall and treating water differently from the broader water-footprint approach. (Springer)
Wong et al., 2021, Water Resources Research	California Central Valley crops	Almonds accounted for 22.2% of crop consumptive water use in water year 2014	Shows that the issue isn't merely water per almond: almonds were the largest single crop category in Central Valley consumptive water use in that year. (AGU Journals)
Stewart et al., 2011, California Agriculture	Almond irrigation	Deficit irrigation saved about 5 inches of water/year without significant yield reduction	Demonstrates that almond water intensity isn't fixed; irrigation technology and management can materially change it. (California Agriculture)
Li et al./Ren et al., 2023–25, Making AI Less "Thirsty"	GPT-3	~700,000 L direct water for training; ~5.4 million L including broader water footprint	Demonstrates that AI's water footprint includes both inference and the much larger infrastructure/training question. (DOI)
Google, 2025, production Gemini measurement	Gemini text inference	0.26 mL/prompt comprehensive measurement	One of the few production-scale measurements rather than a theoretical estimate. (Google Cloud)
Green, 2026, The Hidden Thirst of AI	GPT-4o and Gemini	GPT-4o: ~0.75 mL total in its reference scenario; published ChatGPT figure 0.322 mL	Illustrates how including electricity-generation water can substantially increase the apparent AI footprint. (MDPI)

On the computing consumption front, some might argue the water consumed to build and operate data centers has to be counted, as well as the water to produce servers, connectors and cables, for example. 

So one might plausibly argue that producing one almond is equivalent to as few as 350 prompts or as many as 23,000. 

AI water estimate	Equivalent to one almond
0.26 mL — Google's measured Gemini prompt	~23,000 prompts
0.32 mL — OpenAI/Altman ChatGPT estimate	~18,750 prompts
0.75 mL — GPT-4o broader estimate including indirect water	~8,000 prompts
6 mL — higher-end estimate for some AI tasks	~1,000 prompts
17 mL — high-end GPT-4o estimate in a recent study	~350 prompts

That provides some perspective on the “AI data centers use too much water” claim, but also highlights the water intensity of almond growing. Compared to growing lettuce, almonds require 6690 percent more water, for example.

Crop	Green water	Blue water	Gray water	Total	Almonds vs. crop
Almonds, shelled	9,264 L	3,816 L	3,015 L	16,095 L	—
Pistachios	3,095	7,602	666	11,363 L	42% more
Cashews	12,853	921	444	14,218 L	13% more
Walnuts, shelled	5,293	2,451	1,536	9,280 L	73% more
Lentils	4,324	489	1,060	5,874 L	174% more
Dry beans	3,945	125	983	5,053 L	218% more
Sorghum	2,857	103	87	3,048 L	428% more
Soybeans	2,037	70	37	2,145 L	650% more
Wheat	1,277	342	207	1,827 L	881% more
Rice, paddy	1,146	341	187	1,673 L	962% more
Maize/corn	947	81	194	1,222 L	1,217% more
Potatoes	191	33	63	287 L	5,508% more
Tomatoes	108	63	43	214 L	7,521% more
Sugar cane	139	57	13	210 L	7,665% more
Lettuce	133	28	77	237 L	6,690% more

Roughly the same argument, though, can be made about agricultural water use and computing water use. Nationally, farms consumed nearly 30 trillion gallons of water for crop irrigation in 2020, the most recent data available from the U.S. Geological Survey. Meanwhile, data centers nationwide used about 228 billion gallons of water in 2023.

Yes, data centers consume water, though other alternatives provide ways of consuming less water. But data centers consume vastly less water than agriculture: hundreds to many thousands of times more water, in fact. 

Compared to other necessary uses, data center water use is miniscule, in fact. 

source: Axios 

In the intermountain U.S. west, all industry combined consumes five percent to 10 percent of water, while agriculture consumes 70 percent to 80 percent, for example.  

Sector	Share of Water Consumption (Typical Western Basin)
Agriculture	70–80%
Municipal	10–20%
Industry	5–10%